Exploring Web Scale Language models for Search Query Processing
Proceedings of the 19th International World Wide Web Conference (WWW'2010), Raleigh, NC |
It has been widely observed that search queries are composed in a very diﬀerent style from that of the body or the title of a document. Many techniques explicitly accounting for this language style discrepancy have shown promising results for information retrieval, yet a large scale analysis on the extent of the language diﬀerences has been lacking. In this paper, we present an extensive study on this issue by examining the language model properties of search queries and the three text streams associated with each web document: the body, the title, and the anchor text. Our information theoretical analysis shows that queries seem to be composed in a way most similar to how authors summarize documents in anchor texts or titles, oﬀering a quantitative explanation to the observations in past work. We apply these web scale n-gram language models to three search query processing (SQP) tasks: query spelling correction, query bracketing and long query segmentation. By controlling the size and the order of diﬀerent language models, we ﬁnd that the perplexity metric to be a good accuracy indicator for these query processing tasks. We show that using smoothed language models yields signiﬁcant accuracy gains for query bracketing for instance, compared to using web counts as in the literature. We also demonstrate that applying web-scale language models can have marked accuracy advantage over smaller ones.